That's not true for NLU at least. It is on par with 2018's RoBERTa on GLUE, many larger and advanced language models came after.

It is still great work though, a robust masking representation architecture that works across modalities.

They mention that on the article:

> We apply data2vec separately to speech, images and text and it outperformed the previous best single-purpose algorithms for computer vision and speech and it is competitive on NLP tasks.